The Physical Ceiling: What No Model Can Recover From Your Sensor
Mostafa Dhouib··7 min readWhatever your sensor and that first little circuit throw away is gone. You cannot get it back later with software. A photo that comes out blurry does not become sharp with editing, and the fancy tools that appear to sharpen it are inventing plausible detail rather than recovering real detail, which is exactly the distinction that matters when the output drives a decision about a person.
The Physical Ceiling: What No Model Can Recover From Your Sensor
The short answer. Whatever your sensor and its first circuit throw away is gone. No amount of processing gets it back, because the information does not exist anywhere downstream to be recovered. A photo that comes out blurry does not become sharp with editing. Modern tools can make it look sharp, and that is a different operation: they are inventing plausible detail consistent with what was captured, not recovering the detail that was lost. For a picture of a holiday that distinction is cosmetic. For a measurement that drives a decision about a person, it is the whole thing.
There is a ceiling on every measurement system, it is set in the first inch of the signal chain, and almost every roadmap I see is written as though it does not exist.
Where the ceiling is set
Follow the chain. The physical quantity exists at the body. An electrode or transducer couples to it. A first analog stage amplifies and filters. A converter samples it. Everything after that is arithmetic.
Every stage before the converter can destroy information permanently. Attenuation that puts the signal under the noise floor. A filter that removes the band the information lived in. Saturation that clips a peak. A sampling rate that folds high-frequency content into your signal band, where it is now mathematically indistinguishable from real content. Quantisation that discards resolution below one step.
After the converter, nothing can add information. Processing can reorganise it, denoise it in a statistical sense, and present it more usefully. It cannot create what was not captured.
That asymmetry is the ceiling: the information content of your data is fixed at the point of digitisation, and everything downstream operates on what survived.
The photo, and the sharper version of the photo
The intuition everyone has is the blurry photo, and it is right as far as it goes. Snap a blurry picture and no amount of editing makes it genuinely sharp.
The objection everyone raises is that modern tools clearly do sharpen blurry photos, and the results are impressive.
Both are true, and the reconciliation is the important part. Those tools are not recovering the lost detail. They are generating detail that is plausible given what was captured and given everything else they have seen. The output is a well-informed guess, constrained by the measurement but not determined by it.
For a photograph, an invented plausible detail is usually fine and often desirable. For a measurement, it is a specific and dangerous failure, because the output is no longer a measurement. It is a prediction wearing a measurement's clothes, and every downstream consumer will treat it as the former.
The question to ask of any processing stage that appears to improve a signal: is this constrained by information in the measurement, or is it filling in from a prior? Both are legitimate operations and they carry completely different guarantees. A model trained to reconstruct a physiological waveform from a degraded one will produce something that looks like a physiological waveform, because that is what it was trained to produce, whether or not the underlying event was present in the data.
The two failures this causes
The roadmap that plans to fix it in software. A device ships with a placement that loses most of the signal, on the understanding that the algorithms team will make up the difference. They cannot, and the discovery arrives at month nine when the hardware is locked and the schedule is public. This is why the sensor placement decision is an algorithm decision, made in month two by people who do not work on algorithms.
The model that learns to hallucinate the measurement. Train on degraded inputs paired with good labels and the model will learn a mapping. It will be right often, because physiology is structured and priors are informative. It will be confidently wrong exactly when the case is unusual, which is the case you built the device to catch.
How to find your ceiling before you commit
Four steps, and none needs a finished product.
Measure what arrives, not what you hope for. Capture raw signal at the intended location, on real people, under motion, with your actual front end. Compare it against the same measurement at the best available location. The ratio between those is your ceiling, and it is a measurement rather than an assumption.
Establish what the task needs. Independently of your device, determine what signal quality the downstream decision actually requires. Often this can be answered by degrading good data until performance falls off, which is a day's work and gives you a threshold.
Compare the two numbers. If what arrives clears what the task needs, you have headroom and the rest is engineering. If it does not, no processing will close the gap, and you are choosing between moving the sensor, changing the front end, or changing the claim.
Repeat under the ugly conditions. Motion, poor contact, sweat, dry electrodes, cold hands, the wearer who does not follow instructions. The ceiling under good conditions is not the ceiling that governs your product.
When a model is the right answer
This is not an argument against models, and the honest version has to include the other side.
Sometimes the information genuinely is present in the measurement and is simply hard to extract, buried under structured interference, spread across channels, or in a nonlinear relationship with the thing you want. That is exactly what a model is good at, and reaching for one is correct.
The distinction is whether you are asking the model to extract information that survived digitisation or to supply information that did not. The first is signal processing with more capacity. The second is generation, and it will look identical in a demo.
The test is the one above: measure what arrives, establish what the task needs, and compare. If the information is there, a model may be the best tool for getting it out. If it is not, the model will produce plausible output anyway, and you will not be able to tell from the output which situation you are in.
The sentence worth keeping
Whatever the first circuit loses is gone. Everything after it is arithmetic on the survivors, and arithmetic does not create evidence.
That is not a counsel of despair. It is the reason the cheapest engineering decision available in any measurement system is the one made in the first inch, and the reason it is worth taking seriously in month two rather than month nine.
FAQ
Can software recover a weak or degraded biosignal? No. Information destroyed before digitisation does not exist downstream, and processing operates only on what survived. Attenuation below the noise floor, filtering out the relevant band, clipping, and aliasing are all permanent, and no algorithm reverses them.
But AI tools clearly sharpen blurry photos. Why doesn't that work here? Because they are not recovering lost detail, they are generating plausible detail consistent with what was captured and with what they were trained on. For a photo that is usually fine. For a measurement it converts the output from a measurement into a prediction that will be treated as a measurement.
How do I know whether my sensor placement has enough signal? Measure what actually arrives at the intended location, on real people, under motion, with your real front end, and compare it against the best available location. Separately determine what quality the downstream task needs by degrading good data until performance drops. Compare the two numbers.
When is a model the right answer for a weak signal? When the information survived digitisation but is hard to extract, buried under structured interference, spread across channels, or nonlinearly related to what you want. That is signal processing with more capacity, and it is a good use. Asking a model to supply information that was never captured is generation, and it looks identical in a demo.
Why does this decision have to be made so early? Because it is set by the sensor and its first circuit, which are fixed by the mechanical and industrial design long before anyone looks at a signal. By the time the algorithms team discovers the ceiling, the hardware is locked and the schedule is public.
Tell us the system, the stakes, and the date that matters. You get a straight technical reply from the person who would lead the work, within 24 hours.
Bring us the program